%----------------------------------------------------------------------------

NATURAL LANGUAGE UNDERSTANDING (NLU) -
LEXICON, SYNTAX and SEMANTICS 

%----------------------------------------------------------------------------

SEMANTIC GRAMMARS:

Semantic grammars encode the semantics in the grammar itself.
They represent the meaning of the terms procedurally, not declaratively 
(descriptively and explicitly).

Example: SHRDLU (Winograd) and LADDER [Hendrix]

Semantic grammar for a query language for querying a database
containing information about naval ships.


S ---> what is the SHIP-PROPERTY of SHIP?
SHIP-PROPERTY ---> the SHIP-PROP | SHIP PROP
SHIP-PROP ---> speed | length | beam | type
SHIP ---> SHIP-NAME | the fastest SHIP2 the biggest SHIP2 | SHIP2
SHIP-NAME ---> Kennedy | Kitty Hawk | ...
SHIP2 ---> COUNTRYS SHIP3 | SHIP3
SHIP3 ---> SHIPTYPE LOC | SHIPTYPE LOC
SHIPTYPE ---> carrier | submarine
COUNTRYS ---> American | British
LOC ---> in the Pacific | in the Atlantic


In this grammar, words in captial letters are semantic categories;
other words are syntactic place markers.

This grammer can be used to parse sentences such as
``what is the speed of the fastest American carrier in the Pacific?''

Captures semantics in an limited sense --- procedural semantics.
Can be implemented on an ATN.

The grammar is limited to a very narrow domain.
As the domain size increases, the number of grammar rules
increases rapidly.


%----------------------------------------------------------------------------

CASE GRAMMARS: [Filmore ~1965]

The two sentences 
1. Mary hit Bill
2. Bill was hit by Mary
give two very different parse trees.

However, their ``meaning'' (in terms of objects and relations
in the real world) is pretty much the same.

Case grammars represent the meaning of sentences around
actions such as ``hit''.

Action: hit
	Agent: Mary
	Dative: Bill

Now, the same representation expresses the meaning of both sentences.
	
Of course, this language needs to be augmented for representing more
complex sentences such as

Mary hit Bill in the zoo;
Mary hit Bill in the afternoon; etc.

There is no agreement on what set of semantic primitives are needed.
Some possible primitives are

Agent: (animate) instigator of action
Dative: (animate) entity affected by action
Object: entity acted upon
Instrument: (inanimate) cause of event or object used in causing the event
Factitive: (inanimate) object or (animate) being resulting from event
Locative: place of event
Source: place from which something moves
Goal: place to which something moves
Beneficiary: (animate) being on whose behalf the event occurred
Time: time of event
...
...

The parsing is done as follows.
First, the verb in the sentence is located.
Then, the verb is used as an index to select the coresponding action
schema from memory.
Next, the action schema generates expectations about other phrases
and terms in the sentences, e.g., the action schema corresponding
to the verb ``hit'' generates expectations about an agent, a dative,
and so on.
Finally, these expectations help to parse the sentence in terms of the
semantic categories indicated by the action schema.

Again, ATNs provide a good parser for case grammars.

Advantages:
Captures semantics declaratively (descriptively and explicitly).
Generates expectations (once the verb of the sentence is located,
it can be used to predict noun phrases in the sentence etc.)

Drawbacks: 
Schema selection problem --- because of the large number
of verbs in a natural language, the memory may contain a 
correspondingly large number of action schemas.
Expectation generation problem --- because of syntactic issues,
e.g., active vs. passive voice, a large number of rules for
generating expectations may be needed.

%----------------------------------------------------------------------------

Copyright (c) Ashok Goel
Associate Professor, College of Computing 
Georgia Institute of Technology, Atlanta, Georgia 30332
Email: goel@cc.gatech.edu

%----------------------------------------------------------------------------